train the model. we run the pipeline.

◉ the mechanism · what actually happens

One request, walked end to end — every step on your infrastructure, against your own models, with a human on the gate.

1. 01
   ChatMode

   You state the goal in plain language — “run the training job, score the eval, register the model.” ChatMode plans the work and dispatches its built-in agents — planning, data-query, validation, synthesis — to decompose it into pull, analyze, act, assemble. No prompt engineering, no scripts.

2. 02
   SmartModelRouter

   Each step is routed to a model that fits it — a fast model to tag and sort data, a long-context model to read an eval report or a long training log, a strong reasoner to judge whether a run regressed or improved — across the providers you register. Bring your own models and run them on your own infrastructure; your training data, your weights, and your eval sets stays inside your network.

3. 03
   MCP tools · OBO credentials

   Agents reach your feature store, object storage, training clusters, and the model registry as MCP tools, each call running under your own identity. The platform forwards your scoped credentials per call — no shared token, no pooled service account, no secret pasted into a prompt. Every call is logged.

4. 04
   Embeddings · vector stores

   The embeddings that power your retrieval run on your own infrastructure too — throughput tracked per provider and model, indexed into the vector stores you register, so the representations of your data never leave your network. You see what’s being embedded and where it lands, not a black-box hosted index.

5. 05
   CodeMode · Code Execution

   The glue work runs against your real code. CodeMode is an enterprise-platform capability that emits telemetry to your own monitoring inside your boundary — never your data, and nothing ever leaves your network; for one-shot transforms, Code Execution will write a one-shot data-prep or eval-scoring tool, run it sandboxed, and discard it. Nothing persists, and it re-runs as the inputs change.

6. 06
   AgenticWorkflows

   The train → evaluate → register flow becomes a flow of named agents — repeatable, auditable, on rails. And there’s more behind sign-up: data-quality sweeps, hyperparameter search orchestration, and the drift-and-retraining triggers — the many ways the platform runs this, with more revealed after you sign in.

7. 07
   HITL approval

   Nothing acts until you say so. The human-in-the-loop gate is real architecture, not a setting — any step that would promote a model or kick off a training run stops and waits for a person; if no one approves, it times out and is denied. You review the plan and the scope, and approve before anything happens.

8. 08
   Code Mode · build the pipeline tool, don’t wait for it

   The data-prep tool, the eval harness, the registry hook you keep meaning to write — Code Mode drafts it on your stack and proves it with tests, gated by your review before it runs against your data. You stay on the modeling and the eval judgment; the agent builds the pipeline plumbing you’d otherwise queue for months, so a small ML team ships at a bigger one’s velocity.

9. 09
   Audit trail · DLP · RBAC

   Every model call, every tool call, every approval is written to an append-only audit log — once a decision is recorded it’s frozen, so the trail is tamper-evident. DLP keeps sensitive material inside your network, and role-based access scopes who can do what. It’s a record you can stand behind.

The toil ran itself, it acted only on your nod, and the record is already written — and that’s one flow. Data-quality sweeps, hyperparameter search orchestration, and the drift-and-retraining triggers are waiting behind sign-up.